A new study published in Environmental Advances offers a strikingly detailed view of this process. Researchers led by Waluyo Yogo Utomo combined machine learning with an unusually dense network of field measurements to map groundwater levels across an entire peat hydrological unit on Rupat Island at a resolution of 10 meters, updated every two weeks, spanning 2019 to 2025. They then converted those water table maps into carbon dioxide emission estimates using a newly fitted response curve built from nearly a hundred chamber measurements across Indonesian peatlands. The approach, the authors argue, could replace the static, land-use-based emission factors that currently dominate national greenhouse gas accounting with something far more responsive to climate reality.

The scale of the data collection is itself remarkable. The team drew on 134 automatic loggers and manual dipwells installed across the Rupat Island peat hydrological unit, a landscape of roughly 118,500 hectares containing two peat domes up to 12 meters deep. To anchor the model in permanently saturated zones, they added 40 dummy points along estuarine and coastal margins. In total, the dataset comprised 16,496 biweekly water level records. These were paired with an arsenal of satellite-derived predictors: rainfall from CHIRPS, evapotranspiration from MODIS, radar backscatter from Sentinel-1 as a proxy for soil moisture, and a suite of static variables including peat thickness, lidar-derived elevation, slope, flow accumulation, topographic wetness, and distances to canals, rivers, and the coast.

The algorithm at the heart of the study is Extreme Gradient Boosting, or XGBoost, a tree-based machine learning method prized for its speed and accuracy on tabular data. After screening 47 candidate predictors down to 33 to eliminate multicollinearity, the team trained the model on 70 percent of the records and validated it on the remaining 30 percent. The results were strong: the calibrated model explained 79 percent of the variance in water level with a root mean square error of 7.5 centimeters, while independent validation achieved 73 percent explanatory power and an error of 8.3 centimeters. Repeated Monte Carlo cross-validation confirmed that the performance was stable rather than a statistical fluke.

Perhaps most illuminating is what the model learned about its own decisions. Using SHAP analysis, a technique borrowed from game theory that assigns each predictor an additive contribution to every prediction, the researchers found that filled surface elevation was the single most important driver of water table depth, followed by distance to the coast and antecedent climate conditions. Years with pronounced drought, notably 2020 and 2022, also carried substantial weight. In other words, the shape of the land and the memory of past rainfall matter more for peatland hydrology than any single snapshot of weather, a finding that echoes the physical intuition of peat scientists but had rarely been quantified at this scale.

The mapped water tables revealed a landscape under quiet stress. Average water levels across the unit sat at about 36 centimeters below the surface, within Indonesian regulatory limits, but spatial differences reached 30 to 60 centimeters. Drained plantation areas on the coastal backswamps plunged to depths of 45 to 60 centimeters during dry years, while undrained forest remnants, squeezed into small patches at the domes’ centers, showed water tables of 30 to 40 centimeters, deeper than comparable undrained forests elsewhere in Sumatra and Kalimantan. The authors suggest these forest islands are hydrologically hostage to the drained plantations surrounding them, a sobering illustration of how drainage infrastructure reshapes entire landscapes.

To translate water depth into carbon, the team systematically reviewed 32 closed-chamber studies across Indonesian peatlands, distilling 97 observations of soil respiration and heterotrophic respiration, the component that represents genuine peat decomposition rather than root activity. Fitting a Gompertz, or sigmoidal, function to these data produced a markedly better description of the emission-water table relationship than the linear models used in earlier compilations, with explanatory power of 22 to 24 percent versus just 4 to 5 percent previously. The curve captures a crucial nonlinearity: emissions respond sharply to water table changes between 0 and 60 centimeters below the surface, but flatten out when peat is fully inundated or desiccated beyond microbial tolerance.

Upscaling the response curve across the biweekly water maps yielded striking totals. The Rupat Island unit emitted an average of 52.1 megagrams of carbon dioxide per hectare per year as soil respiration, of which about 42 megagrams came from heterotrophic peat oxidation, a ratio of roughly 0.80 consistent with prior studies. Across the whole landscape, that amounts to 4.44 million tonnes of carbon dioxide per year from soil respiration and 3.55 million tonnes from peat decomposition alone, or 31.1 and 24.8 million tonnes cumulatively over the study period. Emissions peaked during the dry early months of 2020, when rainfall hit long-term minima and water tables plunged, and simulations showed that a dramatic drawdown from 0 to 80 centimeters below the surface could inflate emissions by nearly 200 percent.

The asymmetry embedded in the response curve carries a hard policy lesson. A 10-centimeter drop in the water table from 40 to 50 centimeters deep produces a larger emission increase than the reduction achieved by raising it back from 50 to 40 centimeters. Drainage, in other words, is far easier to cause than to undo, and short-term rewetting cannot fully compensate for drying episodes. The study’s estimates also diverged from the static benchmarks: IPCC Tier 1 emission factors produced the lowest heterotrophic emission figures, about 88 percent of the new estimate, while Indonesia’s forest reference level and earlier linear models produced higher ones. The differences widen cumulatively over time, meaning national inventories built on static factors may systematically misjudge both the magnitude and the timing of peat carbon losses.

The authors are candid about limitations. Dipwell networks remain sparse in inaccessible swamp interiors, the chamber studies are unevenly distributed across land uses, and the estimates exclude fire emissions, methane, nitrous oxide, and carbon exported through rivers. Extrapolating the Rupat model nationally would introduce substantial uncertainty, since the algorithm learned site-specific relationships between canals, topography, and climate. Yet the framework points clearly toward a future in which emission factors are dynamic, localized, and sensitive to the El Niño cycles and climate variability that increasingly govern tropical peat hydrology. As monitoring networks expand and the method is validated across other peat hydrological units, it could give Indonesia, and the world, a far sharper instrument for tracking one of the largest and most vulnerable carbon reservoirs on Earth.

Subject of Research: Machine learning-based estimation of groundwater level and carbon dioxide emissions in Indonesian tropical peatland

Article Title: MACHINE LEARNING-AIDED GROUNDWATER LEVEL AND CO 2 EMISSION ESTIMATIONS IN INDONESIAN TROPICAL PEATLAND

Article References: Utomo, W. Y., Aini, F. K., Askary, M., Anwar, S., Tarigan, S. D., & Barus, B. (2026). MACHINE LEARNING-AIDED GROUNDWATER LEVEL AND CO2 EMISSION ESTIMATIONS IN INDONESIAN TROPICAL PEATLAND. Environmental Advances, Article 100762. https://doi.org/10.1016/j.envadv.2026.100762

Image Credits: AI Generated

DOI: 10.1016/j.envadv.2026.100762

Keywords: tropical peatland, groundwater level, carbon dioxide emissions, machine learning, XGBoost, Indonesia, peat decomposition, emission factors, remote sensing, SHAP analysis, Rupat Island, carbon accounting

The study, led by Camilla Citterio with Nicolò Golinucci, Lorenzo Rinaldi, and Matteo Vincenzo Rocco, is the first integrated, cross-paradigm assessment of its kind. Rather than comparing studies within a single methodology, the team aligned emission factors from process-based life cycle assessment databases such as Ecoinvent, environmentally extended input–output tables including EXIOBASE, EMERGING, EORA, GLORIA, and GTAP, primary data from the International Energy Agency, and harmonized reference values from the IPCC, the National Renewable Energy Laboratory, Electricity Maps, the European Commission’s Joint Research Centre, eGRID, and EMBER. To make these sources commensurable, the researchers mapped every technology category onto a shared taxonomy and standardized all greenhouse gas figures to 100-year Global Warming Potential values from the IPCC’s Fifth Assessment Report, then computed national grid intensities for EU countries and the United States using actual generation mixes from the ENTSO-E Transparency Platform and eGRID.

A central contribution of the paper is terminological clarity. The authors distinguish between direct carbon intensity—the operational emissions from generating one kilowatt-hour, equivalent to Scope 1 for a power producer—and life-cycle carbon intensity, which extends the boundary to upstream processes like fuel extraction and plant construction and sometimes downstream impacts such as transmission losses. They also draw the crucial line between attributional factors, which allocate the average emissions of a grid across all consumed electricity, and consequential factors, which estimate which power plant responds to an additional unit of demand. Their analysis focuses on national-scale attributional factors, the common ground for corporate carbon footprints, since consequential modeling requires dedicated tools suited to short-term operational signals like smart charging.

The results reveal a striking asymmetry. For direct operational emissions of fossil fuel plants, the sources show moderate agreement: the spread between the lowest and highest median estimates is 36 percent for coal, 52 percent for oil, and 116 percent for natural gas. Once the boundary expands to the full life cycle, however, consensus collapses. Life-cycle spreads for fossil technologies range from 101 to 631 percent, and for renewables—where emissions are dominated by manufacturing and decommissioning rather than combustion—the choice of data source becomes decisive. The reported life-cycle intensity for hydroelectric power in Ecoinvent and the upper end of the IPCC range substantially exceed every other source, a difference attributed to the inclusion of diverse plant types and fugitive reservoir emissions. For nuclear power, the Electricity Maps factor is less than half the IPCC median and falls below the minimum of the NREL harmonization range.

One discrepancy stands out as a cautionary tale about accounting rules. The IEA reports direct carbon intensities for biomass-fired electricity that can exceed 2,000 grams of CO2 equivalent per kilowatt-hour, while other sources report near-zero direct emissions. The gap stems from a single convention: the IEA counts biogenic CO2 in its factors, contrary to IPCC guidelines that treat such emissions as carbon-neutral within the energy system’s carbon cycle. A differing interpretation of one rule therefore produces results that differ by orders of magnitude—potentially misleading for policy and reporting alike, and a stark illustration that even ‘direct’ emissions are not as unambiguous as they appear.

Beneath the scatter, the researchers uncovered a systematic pattern they call the paradigm effect. EXIOBASE Hybrid, an environmentally extended input–output dataset, consistently yields the highest or among the highest life-cycle intensities for nearly every technology analyzed. For natural gas, the IPCC and NREL medians align at 490 grams per kilowatt-hour and Electricity Maps clusters tightly around 529, while Ecoinvent’s median reaches 598—but EXIOBASE Hybrid’s median soars to 780. For coal, the entire EXIOBASE Hybrid distribution sits above the third quartile of the other major sources. For wind, where all medians fall within a narrow 11 to 23 grams per kilowatt-hour, EXIOBASE Hybrid’s outliers surpass 150. The authors are careful to stress that this is not an error: input–output models capture the full economic web of services and capital goods that process-based inventories systematically truncate, trading technological specificity for completeness.

These technology-level divergences propagate directly to national grids. Comparing computed direct carbon intensities against the IEA reference, the relative difference between minimum and maximum estimates ranged from 72 to 730 percent across countries in 2017, and from 38 to 612 percent in 2023. Monetary-based sources—where emissions per euro are converted to physical units using regional electricity prices—tended to yield lower estimates, and their life-cycle spreads ran roughly 150 percent wider than physical-unit databases on average. EXIOBASE Hybrid stood above the country median in 96 percent of cases and was the single highest estimate for 20 of 28 countries in 2017. The authors emphasize that monetary factors are inherently mediated by electricity prices, making their uncritical use in formal reporting highly problematic.

The stakes become tangible in the paper’s worked example. A 10 megawatt-hour electricity bill in Italy, costing 1,600 euros at 2023 basic prices, produces a maximum total greenhouse gas estimate three times the minimum—a difference of four tonnes of CO2 equivalent for the same consumption, in the same country, in the same year. The methodological choice, in other words, can dominate the result. The researchers also note that corporate practice amplifies the problem: a 2025 study cited in the paper found that 25.8 percent of 10,867 reporting companies used spend-based approaches, and three-quarters of those specifying a method relied on single-region input–output models, implicitly applying domestic emission intensities to imported goods and underestimating the often dirtier footprint of global supply chains.

From these findings the team distills a practical decision framework rather than a single verdict. There is no optimal data source, they conclude—only context-appropriate ones. For basic Scope 2 reporting, primary data from the IEA offers a widely accepted foundation, provided users remain vigilant about biogenic carbon conventions. When a conservative, upper-bound estimate capturing the broadest economic context is desired, input–output databases are defensible, and the authors argue they are the only approach enabling coherent allocation of Scope 3 emissions across an entire value chain. When comparability with product life cycle assessments or Environmental Product Declarations is the goal, process-based sources are the necessary choice. Above all, every reported factor must be accompanied by a clear citation of its source and version—the minimum requirement for reproducibility.

The paper arrives as the GHG Protocol moves toward updating its Scope 2 Guidance with greater emphasis on temporal and geographical granularity, and as the Protocol and ISO work toward a harmonized standard. The authors welcome these developments but caution that they only partially address the core problem of consistency across data sources. Without more reliable primary data, definitive guidance on imports and exports, robust residual-mix methodologies, and systematic documentation, significant discrepancies are likely to persist. Their prescription is a centralized, regularly updated reference dataset applied consistently over time, complemented by primary data where feasible. In a rapidly decarbonizing world—and amid polarized debates over sustainability regulation—the message is blunt: the carbon intensity of electricity is not a fixed fact of nature but a context-dependent figure, and credibility in climate policy depends on transparent sourcing and uncompromising communication of analytical boundaries.

Subject of Research: Methodological comparison of electricity carbon intensity emission factors across life cycle assessment, input–output, and agency data sources

Article Title: Carbon intensity of electricity: a systematic methodological and quantitative review

Article References: Citterio, C., Golinucci, N., Rinaldi, L., & Rocco, M. V. (2026). Carbon intensity of electricity: a systematic methodological and quantitative review. Journal of Industrial Ecology, 30(4), 2173-2187. https://doi.org/10.1007/s44498-026-00148-3

Image Credits: AI Generated

DOI: 10.1007/s44498-026-00148-3

Keywords: carbon intensity, electricity, emission factors, life cycle assessment, input-output analysis, EXIOBASE, Ecoinvent, GHG Protocol, Scope 2 emissions, corporate sustainability reporting, grid emissions, industrial ecology

The study was conducted by Mohamed Boullouz, Mohamed Louay Metougui, and Ngonidzashe Chirinda at the Agricultural Innovation and Technology Transfer Center of Mohammed VI Polytechnic University in Ben Guerir, Morocco. Their experimental site sits in a semiarid zone where wheat, the region’s staple cereal, is grown under drip irrigation, a water-saving technique that delivers water and dissolved nutrients directly to the root zone. Drip systems are often promoted as environmentally friendly because they reduce water waste and can improve nitrogen use efficiency. But as this study demonstrates, the climate accounting of such systems depends heavily on which fertilizer goes into the irrigation lines.

The team set up a field trial comparing four mineral nitrogen sources at an identical application rate of 125 kilograms of nitrogen per hectare: conventional urea, ammonium nitrate, a urease-inhibited urea product marketed as Duramon, and a blend of ammonium nitrate and urea. An unfertilized control plot completed the design. Urea is by far the most widely used nitrogen fertilizer on Earth, prized for its high nitrogen content and low cost, but it must first be hydrolyzed in soil to ammonium before plant roots can take it up. That hydrolysis step creates a burst of ammonium and a localized rise in soil pH, conditions that can supercharge the microbial processes, nitrification and denitrification, that produce nitrous oxide as a byproduct.

Over the growing season, the researchers measured cumulative nitrous oxide fluxes using standard chamber-based gas sampling. The results were striking. Plots fertilized with conventional urea emitted 1.84 kilograms of nitrogen per hectare as nitrous oxide, while ammonium nitrate and the ammonium nitrate plus urea blend each emitted just 1.02 kilograms. The urease-inhibited urea fell in between at 1.36 kilograms. The unfertilized control produced 0.28 kilograms, establishing the background emission against which fertilizer-induced losses are judged. Expressed per unit of applied nitrogen, urea emitted 1.82 times more nitrous oxide than ammonium nitrate, even though grain yields across the fertilized treatments were statistically comparable.

Those numbers translate directly into emission factors, the percentage of applied nitrogen that escapes into the atmosphere as nitrous oxide. The Intergovernmental Panel on Climate Change currently instructs national inventory compilers to assume a default emission factor of 0.5 percent for dry climates, a figure meant to represent the average behavior of all nitrogen inputs in water-limited regions. The Moroccan measurements tell a different story. Urea’s emission factor was 1.24 percent, with a bootstrap 95 percent confidence interval of 1.10 to 1.38 percent, meaning it exceeded the IPCC default in every one of the thousands of statistical resamples the team performed. The urease-inhibited urea came in at 0.86 percent, also consistently above the default. Only the nitrate-containing fertilizers, at 0.59 percent each, bracketed the IPCC value, with confidence intervals spanning the 0.5 percent line.

The mechanistic explanation lies in soil microbial ecology. When urea hits moist soil, the enzyme urease cleaves it into ammonium within days, temporarily suppressing nitrite-oxidizing bacteria and creating conditions under which nitrifier denitrification and incomplete nitrification release disproportionate amounts of nitrous oxide. Ammonium nitrate, by contrast, delivers nitrogen already split between its ammonium and nitrate forms, bypassing the hydrolysis spike and smoothing out the microbial activity that drives emissions. Urease inhibitors such as the one in Duramon slow the hydrolysis step, and the study confirmed that they do reduce cumulative emissions relative to plain urea. However, the inhibitor did not significantly reduce yield-scaled emissions, the amount of nitrous oxide released per ton of grain produced, suggesting the technology moderates but does not eliminate the urea penalty.

The implications reach well beyond a single wheat field in Morocco. National greenhouse gas inventories under the Paris Agreement rely on IPCC emission factors, and for countries without extensive local measurement networks, the Tier 1 dry-climate default of 0.5 percent is the working assumption. If urea-dominated irrigated systems in semiarid regions actually emit at rates closer to 1.2 percent, those inventories are systematically undercounting agricultural nitrous oxide, and the mitigation policies built on them are aiming at the wrong target. Conversely, the finding that nitrate-based fertilizers hover near the default suggests that the 0.5 percent figure may be reasonable for some fertilizer types but misleading as a blanket value. The study’s authors argue that non-stratified emission factors, ones that ignore fertilizer source, risk underestimating emissions precisely in the urea-heavy systems that dominate global fertilizer consumption.

What makes the result especially compelling is the absence of a trade-off. Climate mitigation in agriculture often asks farmers to accept lower yields, higher costs, or more complex management in exchange for reduced emissions. Here, switching from urea to ammonium nitrate or a urea-ammonium nitrate blend cut nitrous oxide losses by roughly 45 percent while grain yields remained statistically indistinguishable across treatments. In a semiarid environment where every kilogram of grain matters for food security, and where drip irrigation has already been adopted to conserve scarce water, fertilizer substitution is an immediately actionable lever. The caveat is practical: ammonium nitrate can be more expensive or harder to obtain in some markets, and regulatory restrictions on nitrate fertilizers exist in certain countries due to misuse concerns. But where supply chains allow, the agronomic case is straightforward.

The study also adds an important data point to a growing body of work questioning how well global emission factors capture the diversity of real-world cropping systems. Recent research has highlighted that legacy effects of past nitrogen applications, sampling frequency, and irrigation management can all bias emission factor estimates, sometimes causing systematic underestimation. The Moroccan trial, conducted under the precise water control of drip irrigation, isolates fertilizer source as the dominant variable and shows it is powerful enough to double emissions at a fixed nitrogen rate. For semiarid wheat belts stretching from North Africa through the Middle East and Central Asia, where urea is the default nitrogen source and irrigation is expanding, the message is clear: the molecule carrying the nitrogen matters as much as the amount applied, and climate-smart fertilizer policy should treat fertilizer choice as a first-order decision rather than an afterthought.

Subject of Research: Nitrous oxide emissions and emission factors from different mineral nitrogen fertilizers in drip-irrigated semiarid wheat

Article Title: Mineral nitrogen source controls N2O emissions and emission factors in drip-irrigated semiarid wheat

Article References: Boullouz, M., Metougui, M. L., & Chirinda, N. (2026). Mineral nitrogen source controls N2O emissions and emission factors in drip-irrigated semiarid wheat. Environmental Monitoring and Assessment, 198(10), Article 1099. https://doi.org/10.1007/s10661-026-15963-1

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15963-1

Keywords: nitrous oxide, urea, ammonium nitrate, urease inhibitor, emission factors, drip irrigation, semiarid wheat, IPCC Tier 1, greenhouse gas mitigation, nitrogen fertilizer, Morocco, soil microbiology

The team applied an emission factor approach, a well-established technique in environmental accounting that multiplies activity data, such as the volume of flue gas treated or the mass of pollutant removed, by coefficients that describe the greenhouse gases released per unit of treatment activity. In 2019, the most recent baseline year in their analysis, the treatment of air pollutants in the global iron and steel industry generated approximately 5.37 billion kilograms of carbon dioxide equivalent. To put that figure in perspective, the researchers note that it is comparable in scale to the greenhouse gas emissions produced by wastewater treatment and waste treatment, two sectors whose secondary climate footprints have already attracted significant scientific and policy attention.

The chemistry behind these emissions is rooted in the end-of-pipe technologies that steelmakers rely on. Desulfurization systems, which remove sulfur dioxide from sintering plant exhaust, typically work by reacting the gas with limestone or lime, a process that releases carbon dioxide both through the chemical decomposition of carbonate and through the energy consumed in producing and processing the sorbent. Selective catalytic reduction systems, used to abate nitrogen oxides, consume energy and in some configurations release nitrous oxide, a greenhouse gas nearly 300 times more potent than carbon dioxide over a century. Fabric filters and electrostatic precipitators, which capture particulate matter, carry smaller but non-negligible energy penalties. When these burdens are aggregated, sulfur dioxide treatment emerges as the single largest contributor to the industry’s treatment-related greenhouse gas footprint, a finding the authors describe as decisive.

The spatial distribution of these emissions is strikingly lopsided. Asia accounted for fully 91 percent of global greenhouse gas emissions from air pollutant treatment in the steel sector, and within that, China alone contributed 76 percent of the worldwide total. This concentration is not an artifact of the accounting method but a direct reflection of where the world’s crude steel is actually made. China produces more than half of global crude steel, and its vast fleet of sintering machines and blast furnaces generates correspondingly enormous volumes of flue gas that must be treated before release. Where pollutant control is most intensive and production most massive, the secondary climate burden follows.

Temporally, the study traces how these emissions have evolved alongside tightening air quality standards. As countries, China foremost among them, imposed progressively stricter limits on sulfur dioxide, nitrogen oxides, and particulate matter, steelmakers responded by retrofitting desulfurization, denitrification, and dust removal systems across their plants. Each retrofit reduced the pollutant escaping the stack but increased the energy and material inputs consumed by the treatment train itself. The result is a structural dynamic in which air quality gains and climate costs rise together, unless the treatment technologies or the underlying production processes change.

Two factors, the researchers conclude, dominate the magnitude of these emissions: the industrial production structure and the choice of terminal treatment technology. Production structure matters because the pollutant load entering treatment equipment is determined upstream, by how the steel is made. Integrated blast furnace-basic oxygen furnace routes, which dominate in Asia, generate far larger volumes of sulfur-bearing sintering exhaust than electric arc furnace routes built on scrap recycling. Treatment technology matters because different desulfurization and denitrification systems carry different energy and sorbent intensities per unit of pollutant removed. The study’s scenario modeling exploits these two levers to explore what the future might hold.

Under a baseline scenario that extends current trends in production and pollution control deployment, the team projects that greenhouse gas emissions from air pollutant treatment in the global steel industry would climb to roughly 11 billion kilograms of carbon dioxide equivalent by 2050, a doubling relative to the 2019 level. That trajectory would mean that every ton of pollutant abated increasingly comes bundled with a growing carbon bill, quietly eroding some of the net climate benefit of pollution control and complicating national carbon budgets that have traditionally ignored this emission category.

The more encouraging news lies in the mitigation scenarios. The analysis finds that significant reductions in treatment-related greenhouse gases can be achieved by adjusting the production structure itself, principally by shifting from ore-based integrated steelmaking toward higher scrap utilization and electric arc furnace production. Because scrap-based routes generate far less sulfur dioxide at the source, they require less limestone-based desulfurization, and the carbon savings cascade through the entire treatment chain. In other words, the most effective way to decarbonize the cleanup is to produce less pollution in the first place, a conclusion that aligns the air quality agenda with the climate agenda rather than placing them in tension.

The study’s authors argue that their quantification fills a genuine blind spot. Global steel decarbonization roadmaps, including prominent plant-by-plant analyses published in recent years, have concentrated on process emissions from iron reduction and energy use, while the emissions embedded in environmental control equipment have gone largely uncounted. By demonstrating that these secondary emissions are comparable in scale to those of the wastewater and waste treatment sectors, the research gives policymakers a concrete reason to incorporate them into carbon accounting frameworks and to weigh the full life-cycle footprint of pollution control retrofits, particularly in regions planning massive new ultra-low emission programs. As the world pushes simultaneously toward cleaner air and a stable climate, the steel industry’s smokestack scrubbers are a reminder that in environmental systems, nothing is ever truly free, and that the smartest strategies are those that shrink problems at their source rather than paying repeatedly to manage their symptoms.

Subject of Research: Greenhouse gas emissions generated by air pollutant treatment in the global iron and steel industry

Article Title: Spatiotemporal patterns and scenario prediction of greenhouse gas emissions from air pollutants treatment in global iron and steel industry

Article References: Spatiotemporal patterns and scenario prediction of greenhouse gas emissions from air pollutants treatment in global iron and steel industry. (n.d.). https://doi.org/10.1007/s11783-026-2283-9

Image Credits: AI Generated

DOI: 10.1007/s11783-026-2283-9

Keywords: greenhouse gas emissions, iron and steel industry, air pollutant treatment, sulfur dioxide, flue gas desulfurization, emission factors, spatiotemporal analysis, scenario prediction, steel decarbonization, China steel production, ultra-low emission standards, scrap-based steelmaking

The research, led by Andrew Dumit and colleagues at Watershed Technology Inc. in San Francisco, addresses a problem that distinguishes LCI mapping from most standard machine learning benchmarks. In supervised classification, a misclassified item typically produces some anomalous signal that can be detected downstream. In LCI database mapping, by contrast, an incorrect mapping produces no such anomaly. A wrongly matched dataset will dutifully return an emissions number, and that number will look perfectly plausible in a report. The only reliable way to catch such errors has traditionally been item-level expert review, which is prohibitively expensive at the scale of modern corporate procurement data, where organizations may need to map hundreds of thousands of line items. The result has been an industry-wide reliance on category-average emission factors, which sacrifice the resolution that activity-based accounting is supposed to provide.

The team’s solution is a deliberately structured two-agent system that separates the act of proxy selection from the act of quality assessment through what the authors call an information barrier. The first agent, the mapper, proposes LCI database matches using iterative, tool-augmented retrieval, searching the database and refining its candidates much as a human practitioner would. The second agent, the judge, is architecturally prevented from seeing anything the mapper did. It observes only the original input item, the proposed activity, and the activity’s metadata, and then scores the proposed mapping along two independent dimensions: emissions similarity and material similarity. This enforced separation is not an implementation convenience but a methodological choice, designed to prevent the judge from inheriting the mapper’s biases or rationalizing its choices after the fact.

The performance gains reported in the study are striking. On an evaluation set of 1,039 items spanning seven product categories, the mapper achieved 90.7 percent defensible accuracy — defined as the share of items mapped to an option that a domain expert would approve — with zero abstentions. By comparison, retrieval-based baselines reached only 19 to 43 percent, and prior automated systems, while sometimes avoiding outright errors, abstained on 70 to 73 percent of items, effectively punting the hard decisions back to humans. Because the mapper always commits to an answer, every procurement line item receives a concrete lifecycle-based emission factor rather than a coarse category average, which is precisely what item-level carbon accounting requires.

Even more consequential is what the judge component accomplishes. Because proxy errors are silent, the practical value of any automated mapping system depends on how well it can tell its own confident successes from its quiet failures. At a simulated expert review budget of 20 percent, the judge captured 67 percent of all mapping errors, compared with only 37 to 40 percent for heuristic baselines. When the analysis focused on severe errors — cases where the chosen proxy’s emissions deviated by more than 100 percent from the correct value — the judge caught 74 percent of them. This means that organizations deploying the system can concentrate their scarce expert review capacity on exactly the items where a wrong proxy would most distort the reported footprint, rather than sampling randomly or reviewing everything.

The information barrier also yields a property that is rare in applied carbon accounting tools: calibration. Because the judge never sees the mapper’s reasoning, its quality scores function as an independent audit of each mapping, and these scores turn out to be well calibrated against actual correctness. In practical terms, the system can auto-accept 30 percent of its mappings while incurring an error rate of just 0.3 percent on that auto-accepted subset. This selectivity framework connects the work to a broader literature on selective classification and learning to defer to experts, in which models must know not only how to predict but when their predictions deserve trust. The authors note that large language models are often overconfident and biased self-evaluators when asked to grade their own outputs, which strengthens the case for a structurally independent assessor rather than self-reflection.

The technical architecture draws on several strands of recent research. The mapper’s iterative retrieval approach reflects agentic patterns such as ReAct, in which a language model interleaves reasoning steps with tool calls to ground its decisions in external data — in this case, the LCI database itself. The judge’s evaluation role builds on work on LLM-as-a-judge methods, while its scoring dimensions echo established lifecycle inventory practice, notably the data quality indicators introduced by Weidema and Wesnæs in the 1990s and subsequent proxy selection methodologies for choosing the most appropriate LCI dataset. Earlier machine learning applications in life cycle assessment largely focused on classification or regression tasks; the present work differs by treating the mapping problem as one of defensible proxy selection with explicit, auditable quality control, consistent with the requirements of ISO 14044.

The evaluation combined proprietary internal data from Watershed’s technical assessments with a public benchmark. A reproducible subset of 275 items from the Amazon Parakeet dataset of emission factor recommendations has been released, allowing outside researchers to compare their own systems on identical ground truth. The remainder of the evaluation data derives from proprietary procurement records and cannot be published, and the system code itself is proprietary to Watershed, whose carbon accounting products the company commercializes. All authors are Watershed employees, and the paper states that no external funding was received. These disclosures situate the work within a growing wave of industry-led research into AI-assisted sustainability measurement, alongside recent efforts to apply generative AI to product carbon footprint estimation and LCA data quality assessment.

The broader implications reach well beyond one company’s product pipeline. Accurate Scope 3 accounting has become a central battleground in corporate climate accountability, as regulators, investors, and voluntary disclosure frameworks increasingly demand granular, activity-based figures rather than spend-based approximations. By demonstrating that automated mapping can approach expert quality while providing calibrated signals for where human judgment is still needed, the study sketches a plausible division of labor between machines and domain experts at a scale that neither could achieve alone. If such systems mature, the long-standing trade-off between the resolution of a carbon footprint and the cost of producing it may finally begin to loosen, moving organizations from category averages toward genuinely item-level transparency across their supply chains.

The study arrives at a moment when the infrastructure for lifecycle inventory data itself is expanding. Databases such as ecoinvent, which releases documented change reports with each version update, and newer entrants like China’s HiQLCD, continue to grow in geographic and sectoral coverage, yet the fundamental mismatch between what companies purchase and what databases contain persists. This is why proxy selection has long been recognized as a methodological challenge in its own right, with earlier work proposing structured selection methodologies and expert elicitation to patch inventory gaps for specific product categories such as laundry detergents.

The reporting context also matters. Under the Greenhouse Gas Protocol’s Corporate Value Chain standard and its technical guidance, companies are expected to prioritize the Scope 3 categories most relevant to their sector, and sector-specific technical notes from disclosure platforms such as CDP reinforce that purchased goods and services rank highest in relevance for most industries. Spend-based estimation, which multiplies procurement spending by sector-average emission factors, satisfies reporting requirements cheaply but obscures the physical processes actually driving emissions, limiting the value of the resulting figures for procurement decisions and supplier engagement.

One caution raised in the machine learning literature concerns correlated errors: when multiple automated mappings fail in similar ways, headline accuracy figures can mask systematic biases within particular product categories or database regions. The authors’ emphasis on item-level expert annotation of defensible mappings, supported by a dedicated sustainability data advisory team, reflects an awareness that evaluation quality ultimately bounds the trustworthiness of any automated accounting pipeline built upon it.

Subject of Research: Quality-aware automated mapping of procurement items to lifecycle inventory databases for Scope 3 carbon accounting using a two-agent AI system

Article Title: Quality-aware automation for LCI database mapping

Article References: Dumit, A., Rao, K., Ulissi, S., Watson, S., Feintzeig, J., Joyce, P. J., & Bao, S. (2026). Quality-aware automation for LCI database mapping. Journal of Industrial Ecology. https://doi.org/10.1007/s44498-026-00157-2

Image Credits: AI Generated

DOI: 10.1007/s44498-026-00157-2

Keywords: life cycle assessment, Scope 3 emissions, LCI database mapping, carbon accounting, large language models, two-agent AI, proxy selection, quality assessment, selective classification, emission factors, supply chain, machine learning